This week STARK, the German defense company building exactly these systems, raised €500 million led by Sequoia and Founders Fund. I first met STARK’s CEO and Founder, Uwe Horstmann, back at Project A in 2016, years before any of this. Air Street Capital invested because we believe STARK is emerging as a German neoprime: a new prime contractor for unmanned strike systems, built software-first and able to manufacture at the scale a real war demands.

STARK builds autonomous platforms, fast

STARK’s flagship platform, Virtus, is a loitering munition: a drone that flies to a contested area, waits, finds a target, and strikes it. It is not a reconnaissance drone, though it can return and land like one, and it is not a cruise missile. It sits between the two, cheap enough to expend, autonomous enough to find a target on its own, and simple enough to build at scale. Around it STARK is building a family of effectors - Gambit, a man-portable short-range munition, and Cascade, a tube-launched one - alongside Vanta, an unmanned surface vessel that takes the same idea to sea.

Mass you can afford to build

For decades, Western firepower meant a small number of exquisite platforms, each costing millions and far too precious to lose. Ukraine inverted that. The decisive weapon turned out to be the cheap one you can field by the thousand and expend without flinching. A Virtus assembles in roughly ten minutes, and Germany has already put STARK on a framework worth up to €2.8 billion to supply the Bundeswehr, a landmark deal for a new defense company.

Once the weapon is cheap, the constraint moves to manufacturing and a robust supply chain. More than 80% of this round goes into manufacturing and R&D, and STARK is standing up production lines across Germany, the UK, and Ukraine. In a war of attrition, throughput is the moat: the side that can keep replacing what it loses sets the tempo. STARK builds on civilian supply chains and simple lines that stand up in weeks and replicate from city to city, which is both a way to scale and a way to survive being targeted. The defense companies that win this decade will treat the production line as the product.

Lastly, with a growing portfolio of autonomous products, STARK needs software to tie them all together: Minerva. The same software that swarms Virtus in the air coordinates the unmanned boats at sea, navigates when GPS is jammed, and plugs into NATO battle-management. Each platform improves with a software update rather than a new airframe, and every deployment teaches the whole fleet.

Why STARK

Uwe spent a decade as a general partner at Project A, one of Europe’s most active defense investors and an early backer of Quantum Systems, before leaving to run STARK. He has assembled the four branches this company needs at once. To build at scale: Martin Rost, sixteen years at Zalando running a roughly €10 billion unit, and Johannes Schaback, a repeat founder who was CTO of SumUp. To sell into defense: Jan-Patrick Helmsen, former CEO of Rheinmetall's weapons-and-munitions business. To win the politics: Johannes Arlt, until this year a member of the Bundestag's defense committee. And to learn from the war as it is fought: a Ukraine team running an R&D and production hub in Kyiv that turns frontline feedback into design changes.

Regular readers will be familiar with our defense thesis at Air Street Capital. We recently led the Series A for Alta Ares because Europe has to build its own air defense shield. In our letter from this year’s Munich Security Conference, we argued that Europe must move from emergency buying to structural production capacity, and that when a government contracts a domestic firm it confers the industrial gravity that pulls in private capital and localizes supply chains. STARK is what answering that call looks like: sovereign manufacturing, software-defined systems, and a founder building for the war that is actually being fought. The mass that wins the next conflict will be cheap, built at home, and improved in software, and we’re here to stand behind the people making it happen.

A decade ago, the AI community was alight with enthusiasm over the first generative models for images, Generative Adversarial Networks. At the time, the model's outputs looked more like a Microsoft Paint attempt than anything close to photorealism. Ten years later, Latent Diffusion Models, and the scaled-up, improved architectures since built by the team at Black Forest Labs, have ushered in a level of photorealism arguably indistinguishable from reality, were it not for the unreal scenes:

The same trajectory from “pixelated if you’re lucky” to “I can’t tell the difference between reality and AI” is happening in the domain of world models, and Odyssey is leading the frontier.

Last week, Odyssey announced that it has raised a $310M Series B at a $1.45B valuation, led by Natural Capital, with participation from Amazon, GV, AMD Ventures, EQT, IQT, and others. We wrote the first check into Odyssey's seed in late 2023.

On a personal note, I have known both founders far longer than Odyssey has existed: Oliver Cameron from his years building Voyage and then Cruise (the first self-driving car I experienced thanks to him!), and Jeff Hawke from the founding team at Wayve, where I was involved from day 1. I invited Oliver to speak at RAAIS 2023 in London, which created an opportunity for the two of them to spend significant time together in person. They started Odyssey later that year.

This round is a good moment to explain what the team has actually been building. The answer is more interesting than “AI video.”

From videos to worlds

A world model is not a video generator. Sora, Veo, Kling, and their successors take a prompt and render a fixed clip. The output can be beautiful, but the future is locked from the outset of generation. As a consumer of the video, your only mode of interaction is to watch.

By contrast, a world model is closer to a simulator than a renderer. It predicts the next state of an environment from the past and from whatever a participant does next. It accepts input mid-rollout. It holds a persistent state, a memory of the world that can be acted on and changed. Just as next-token prediction enabled language modeling, Odyssey is betting that next-state prediction, at enough scale, forces a model to learn physics, because a model that has not learned physical regularities drifts into nonsense over a long rollout.

The path to frontier world models

Scale is necessary, but not sufficient. The bottleneck is experience, and over the past year Odyssey has attacked it from three directions: 1) making each moment richer, 2) making experience shared and persistent for people and agents, and 3) teaching models to generate their own.

First, richer experience. Most world models are mute, which is not only bizarre to experience but means that a rich information stream is discarded. Sound is where collisions, distance, intent, rhythm, and emotion live. Nothing is in the intellect, as the line goes, that was not first in the senses. As such, Odyssey’s Starchild-1 generates synchronized audio and video in real time, while responding continuously to streaming text, speech, and action. This is the first real-time multimodal world model.

While this sounds logical, the hard part is that audio and video move on different clocks. A small error in one modality can corrupt the other during a long rollout. Starchild-1’s approach is to let each run on its own clock while staying synchronized, turning a bidirectional audio-video foundation model into a causal, real-time world model.

Second, shared experience. Agora-1 is a multi-agent world model that decouples simulation from rendering. One function evolves a shared world state from player actions, while another renders consistent views of that state from independent viewpoints. The result behaves like a game engine with no hand-coded engine underneath: a shared state the model maintains for every participant, which can be edited into new levels while the dynamics hold.

At RAAIS in London on June 12, Jeff Hawke, Odyssey's CTO, ran a live session of an Agora-1-generated GoldenEye death match, with every frame conjured on the fly. Attendees could join the game and play one another in real time.

Beyond games, robots need shared environments before they touch the real world. Agents need places to collide, coordinate, compete, and fail. Simulators need to cover worlds that have never existed. Odyssey is pursuing all of these directions with design partners it will name in time.

Play GoldenEye here!

Third, self-generated experience. PROWL is a reinforcement learning agent rewarded for breaking the world model: freezing a waterfall, losing a crosshair, popping geometry under the camera, ignoring a control input, or collapsing through a hard scene transition. Through active exploration, an agent can therefore improve a world model.

Readers may recall OpenAI’s “Faulty reward functions in the wild” from 2016, where an RL agent steering a boat learned to rack up points by spinning in circles instead of finishing the race. RL agents are unreasonably good at finding the cracks in a system. PROWL points that talent at the world model itself: the agent hunts a weakness, the model trains it away, and the agent comes back for a harder one. It is a way to manufacture the experience these models are short of.

Why Odyssey

When we first invested, we wrote that great research is never enough on its own. The teams that win pair it with execution, product taste, and a feel for where the customer needs to go next.

Oliver Cameron and Jeff Hawke came out of self-driving, where the job is to break an impossible physical AI problem into tractable pieces, simulate the world, make the model robust and generalizable. Odyssey’s shipping cadence since their seed round continues to accelerate: Odyssey-2 Max for scale and physics, Starchild-1 for multimodal grounding, Agora-1 for shared state, and PROWL for closed-loop improvement.

Odyssey is betting that the next leap in machine intelligence comes from systems that build worlds, act inside them, and learn how reality behaves. If that is right, the lead in robotics, agents, and simulation goes to whoever can generate the richest experience fastest. We wrote the first check because we think Oliver, Jeff, and the team are the ones who will.

Join the team!

Guilherme Penedo and Hynek Kydlíček built that data. Over roughly three years at Hugging Face they shipped FineWeb, FineWeb 2, FinePDFs, and FineTranslations, the reference datasets a generation of open models trained on.

The duo have now left to do the same thing for robots with Macrodata. As a FineWeb fan, I’m excited to share that Air Street Capital led their $4M pre-seed alongside a group of angels from the leading AI labs, and today the company comes out of stealth.

Why now

Physical AI is the field’s next scaling story. Jensen Huang calls this “the ChatGPT moment for physical AI,” and the money has followed: robotics drew record venture funding in 2025, and 2026 is on track to dwarf it, with a cluster of companies building robot brains and bodies now carrying multi-billion-dollar valuations - Figure at around $39 billion, Skild around $14 billion, Physical Intelligence reportedly raising near $11 billion. The model side has caught up to the ambition, with vision-language-action models that fold perception and control into one system and world models that let a policy be tested in simulation before it touches hardware.

Every one of those companies needs the same thing to keep scaling: large amounts of well-prepared real-world data. Macrodata does not have to pick which of them wins, it refines the data all of them depend on. And that layer barely exists today. In autonomous driving, models trained on tens of thousands of hours of fleet data already generalize to cities they have never seen; most of robotics has no equivalent.

Physical data is messy in ways text never was: large video files, sensors sampling at different rates, actions and language interleaved, and a dozen incompatible formats with no agreed standard. Teams rebuild brittle scripts every time they swap a robot or a sensor.

What Refiner does

Macrodata’s first product, Refiner, is the tooling for that mess. It is an open-source Python library that reads the formats teams actually use - LeRobot, HDF5 (ALOHA, robomimic, LIBERO), Zarr, MCAP, raw video, Hugging Face datasets - and turns raw episodes into training-ready datasets. You compose a pipeline locally, inspect it in a data viewer built for multimodal data (you cannot cat a video in a terminal), then run the exact same code on managed cloud compute when it is time to process at scale.

Along the way it does the work that lifts policy quality: trimming idle motion, annotating subtasks, tracking what the hands did, and scoring trajectories with reward models, with VLMs run in the loop where a model needs to label or judge the data. You pay for the compute by the second. A pipeline that takes eight minutes on a laptop runs in under a minute on five H100s, for about $0.27.

The same craft, a harder problem

The throughline from FineWeb to Refiner is refinement: the unglamorous work of turning raw capture into the precise signal a model learns from. With text, that meant deduplicating and filtering trillions of tokens until what remained was worth training on. With robots, it means unifying formats, trimming, labeling subtasks, and keeping the demonstrations that teach while dropping the ones that do not. It is the same discipline applied to a harder, less mature, more valuable problem. The business mirrors it: an open-source core that becomes the default way teams handle robot data, and metered cloud compute they run it on.

We backed Guilherme and Hynek because they have done this before: the pair built the open data standard for LLMs. The industry knows that every strong model starts with great data. We think the next generation of robots will too.

Get started with Refiner

Every generation of defense risks building the shield it wishes it had for the last war. France poured concrete into the Maginot Line to stop the invasion it remembered from the First World War. Israel built the Iron Dome to stop barrages of rockets. Both were serious engineering achievements built for threats that moved more slowly than institutions.

That world is gone.

Russia’s war in Ukraine, and Iran’s missile-and-drone onslaught against the UAE and the wider GCC, have exposed the new arithmetic of air defense. Cheap drones, ballistic missiles, cruise missiles, glide bombs, electronic warfare, and massed salvos have changed both the economics and the tempo of the fight. When a cheap drone draws a million-dollar interceptor, the defender may win the intercept and still lose the campaign.

The next shield has to be affordable enough to fire at scale and robust enough to work under jamming. It cannot be static. It must instead evolve with the threat.

That is why Air Street has led the $60M Series A in Alta Ares.

From argument to action

When I wrote in the Financial Times in 2023 that European governments needed to take defense innovation seriously, I meant it as a challenge to governments and the venture industry. Europe had the capital, talent, and technical ambition to build the technologies that safeguard democracy, security, and our way of life. Too often, it chose easier markets, while procurement systems rewarded incumbents built for a slower era.

Since then, the argument has become harder to dismiss. For the last year, I have been looking for the company that could make it real in European air defense: not “AI for defense” slideware, and not a controlled-range demo, but a sovereign, operationally grounded, full-stack company built around a feedback loop from the field.

Alta Ares is that company.

The new air defense stack

Alta Ares is building full-stack, integrated air defense across the entire kill chain: AI-first software, sensors, command-and-control, and effectors built to operate in contested environments against a range of aerial threats.

The company began with software for intelligence, surveillance, and reconnaissance video analysis. Work alongside operators in Ukraine pulled it into the broader air-defense problem: seeing a target, maintaining track, supporting operator decisions, guiding an interceptor, and integrating the result into a system that can actually be fielded.

In many AI applications, failure means a bad answer or a model that needs retraining. In air defense, the object is small, fast, cheap, and often deliberately hard to see. The operator may be tired, cold, under jamming, and making decisions in seconds. A useful AI system cannot live as an analyst beside the workflow. It has to become part of the kill chain itself.

Alta Ares’s products reflect that architecture. Pixel Lock provides onboard computer vision for detection, tracking, and terminal guidance while preserving human control over engagement. Ukrainian drone pilots are already hitting Russian targets from 500km away and as the Financial Times reported, Alta Ares’s terminal guidance software helps interceptors detect and close on Russian drones in the final phase of flight.

Gamma supports autonomous guidance and ISR workflows. X-Lock and Black Bird are both used in the field: X-Lock against short-range one-way attack drone threats, including Shahed-type systems, and Black Bird against faster aerial threats, including cruise missiles and glide bombs.

This isn’t about software grafted onto hardware. Alta Ares develops models, avionics, guidance, operator workflows, and manufacturing around the operational problems faced by the warfighter.

Born from the battlefield

Alta Ares’s most important asset is its feedback loop with the warfighter.

That loop comes from constant in-field deployment. Alta Ares has been active in Ukraine for years and interceptors equipped with Pixel Lock began shooting down Shahed-type drones in 2025. The company has since demonstrated systems with NATO, tested Black Bird in arctic conditions with the Estonian Defense Forces, and is deployed across multiple operational theaters. In rapid succession, Alta Ares has signed large contracts from half a dozen countries across Europe, the Middle East and Asia.

Those milestones matter, but the deeper point is what they make possible. Simulation is useful. Range tests are useful. Neither exposes systems to the full mess of real conflict: electronic warfare, bad weather, changing drone signatures, damaged equipment, uneven training, and the pressure of live operations.

Alta Ares’s data advantage is not a static dataset sitting on a server. It is a relationship with the field: operators use the system, the system sees where it fails, engineers recover the evidence, and the product changes. That is how AI systems become robust in high-consequence environments: not by claiming perfection, but by shortening the distance between failure and adaptation.

That is what the old defense procurement model cannot absorb. As we argued in Bringing Dynamism to European Defense, Europe’s problem has never been talent alone: it is a financial, political, and institutional climate that fails to reward mission-driven defense entrepreneurship. A decade-long program assumes the threat profile will stand still long enough for the program to arrive. In autonomous warfare, that assumption is fatal. The measure of a system is not only how well it works on day one, but how quickly it can be improved for day two.

Why this team

Hadrien Canter’s path into defense technology is not the standard prime-contractor biography. He has deep links with Ukraine for years, and Alta Ares does not feel like a lab looking for a battlefield use case. It feels like a company working backward from the operator: what they can see, what they cannot, how quickly the threat changes, and what a system has to do in the thick of it.

You feel it in the office: the energy, mission, urgency, and seriousness of people building for a live war.

Together with his co-founder Stanislas Walch, Hadrien has recruited software, hardware, government relations and sales talent from Anduril, Helsing, Palantir, Safran, Thales, MBDA, Embraer, and the French Army. They’ve also built an advisory board that includes Philippe Lavigne, former Chief of Staff of the French Air and Space Force and former NATO Supreme Allied Commander Transformation, and General Corentin Lancrenon, a three-star French Army general.

For Air Street, all of these features are critical. We back AI-first companies, but “AI-first” should not mean model-first in isolation. In the physical world, AI advantage usually belongs to the team that owns the loop from data to deployment. Alta Ares has that loop.

Europe must build

Europe now broadly agrees that it needs more defense capability. As I wrote in A letter from the Munich Security Conference, the harder transition is from crisis buying to permanent capacity. The question is whether Europe will build the new capabilities itself, or simply allocate larger budgets to imported systems and slower incumbents. e.

No single company will be Europe’s entire Iron Dome for autonomous air defense. A real adaptive shield will take an ecosystem: sensors, decision systems, interceptors, electronic warfare, command systems, procurement reform, and operators who can field the technology. But every ecosystem needs a company that shows the way.

Europe has written enough policy documents about waking up. It needs companies that jolt us into action.

We believe Alta Ares is that company.

In this piece, I’ll share why a system that connects research, development and clinical decision-making across the full lifecycle of a drug is critical, and why Perceptic is the company to do it.

Can AI accelerate drug discovery?

Drug discovery is expensive, slow and low-hit-rate. We all know this. What is new is that we now live in an era of AI systems that can consume and reason over the multi-modal evidence drug development actually generates - targets, chemistry, biology, clinical, commercial - and across the sprawl of databases, dashboards, notebooks, departments and geographies it lives in.

Look at where the frontier labs are putting their attention. Anthropic recently added Vas Narasimhan, the CEO of Novartis, to its board, and is openly building Claude into a partner for scientific work. OpenAI’s reasoning model just disproved an 80-year-old Erdős conjecture in discrete geometry, verified by external mathematicians on a problem the field had not cracked since 1946. Science - and biology especially - is the pinnacle AI use case, and the labs are not being subtle about wanting it.

A few months ago I wrote Can AI discover new science?, arguing that AI is now an accelerator of discovery, not just an automator of existing work. But drug development is one of the highest-stakes environments that does not come with clean verifiers: no leaderboard, no public benchmark, no “the model proved the lemma” moment. The contributions only count when they are wired into the workflows where billion-dollar, multi-year decisions get made, with every conclusion traceable back to the evidence that produced it. The labs have given us the substrate: frontier models. Pharma now needs the operating system, and this is where Perceptic comes into play.

The Perceptic operating system

Perceptic is the intelligence layer that connects data, decisions and context across the drug lifecycle, so every insight compounds and every decision is made with the full picture. Three AI applications run on one shared architecture:

• Scout triages external assets, including licensing candidates, competitive programs, pipelines, against the customer’s evolving strategy. Evaluation time has gone from a week to an hour in production while asset screening from hundreds per week to thousands in minutes.

• PercepticOS is the intelligence layer above the customer’s internal tools and data. It is where scientific teams test hypotheses, compare internal evidence against external benchmarks, and build a knowledge base that doesn’t restart with every new project.

• Atlas is the clinical data foundation that recapitulates internal and external trial data, providing the substrate everything else stands on. Live deployments have produced a 50-fold increase in clinical data extractions.

Perceptic in practice

A pharma company evaluating a new therapeutic area starts inside PercepticOS, pressure-testing a hypothesis against the evidence base. That triggers Scout to sweep external assets and rank them against the evolving thesis, in minutes instead of weeks. Candidates feed back into PercepticOS with full context, where Atlas surfaces the trial history, benchmarks and endpoint precedents that determine which assets are tractable.

A customer doesn’t buy three products. They deploy AI workers that learn their organization, their tools, their data, their decisions, and the instance gets more valuable every month. Perceptic follows the drug, not the department.

The perfect team, forged at Palantir

Air Street Capital has invested in many AI-first techbio companies including Profluent, Allcyte (acquired by Exscientia), Valence Discovery (acquired by Recursion), all of which are in the business of discovering novel drugs. Perceptic is the opposite bet: the AI-first software that makes drug developers themselves better and faster.

As an investor I am generally of the view that the money is in the drug, not the tools. Perceptic is the exception. The moment has come to bet on the AI-first software layer, and the right team has shown up to build it.

Two ingredients matter, and they are rarely found together. The first is the DNA of shipping production AI into the hardest enterprise environments. Tilman, Martin and Zaki were core contributors to Palantir’s AIP, the suite designed to securely connect frontier AI with an organization’s internal data and operations. They are operators who learned over a decade what it takes to put production AI inside regulated, data-sensitive, multi-stakeholder enterprises. They know where deployments break, and what the six-month security reviews actually ask for.

The second is deep knowledge of pharma workflows themselves. Frontier models are spiky in their capabilities, and the spikes only line up with value when you graft them onto the proprietary workflow knowledge of the end user. The team’s years inside Palantir’s Life Sciences practice mean they understand the nitty-gritty of bending increasingly capable frontier AI systems into the shape pharma R&D actually requires. That is not something you can buy on the open market.

CSL and multiple top-20 pharma companies that we cannot name publicly trusted that thesis enough to deploy Perceptic before the company came out of stealth - the highest-signal proof point any seed-stage company can have.

What comes next

If Perceptic is right, drug development moves from a 15-year linear bench-to-bedside process to one that runs on always-on infrastructure, where every insight from every team is wired into every subsequent decision. The handoffs stop being where information dies: they become where it accelerates.

That is the category that is forming around Perceptic.

Congratulations to Tilman, Martin, Zaki and the whole Perceptic team. We’re proud to be on the journey.

- Nathan

Join the team!

From clip-generators to simulators

For most of the last three years, generative video has meant clip-makers: models that take a prompt and render a fixed-length, fixed-trajectory output video. Veo, Sora, Kling, and their successors have made the visual fidelity of generated video remarkable. They are also fundamentally offline systems. Once generation begins, the future of the clip is locked.

World models are a different flavor of system. First, they predict the next state of an environment given the past, conditioned on what a participant - such as a person, an agent, a robot - does next. Second, they accept streaming input mid-rollout, and the world responds. Third, they hold a persistent, manipulable state, which makes the model a simulator rather than a render.

So what’s changed more recently? Causal video distillation matured (CausVid, Self-Forcing), and bidirectional joint audio-visual foundation models arrived (Ovi, Veo 3). Odyssey has been heads-down expanding these threads into something interactive.

Two new frontier models

Starchild-1 jointly generates audio and video autoregressively at up to 24 fps, while continuously responding to streaming text, speech, or action input. Odyssey frames the case for sound through Aquinas: “nothing is in the intellect that was not first in the senses”. Pretending the world is silent leaves a large amount of signal - physics, dynamics, intent, emotion - out of the model. Audio and video also evolve at very different temporal resolutions, and small errors in either modality compound during long-horizon rollout.

Starchild-1’s answer is a causal distillation pipeline that adapts Ovi, a bidirectional audio-video foundation model, into a real-time autoregressive one, plus an asynchronous KV-cache architecture that lets the two modalities run on their own clocks without losing synchronization. A single model supports four interaction regimes: interactive world exploration, scripted dialogue control, conversational interaction, and narrator-style companionship. The team is candid about what’s left: scene and acoustic identity still drift over long horizons, and quantitative benchmarks for interactive causal audio-video generation don’t yet exist.

Agora-1 matches up to four players into a shared deathmatch - built on GoldenEye, a game many on the Odyssey team grew up playing (as did I) - and every frame each player sees is generated by Agora-1 while the model maintains a shared world state across all participants.

Prior multi-agent work has tried three paths. Multiverse concatenated player views into a single “split-screen” world. Solaris stacked agents along the sequence dimension of one autoregressive transformer - more robust, but context grows with the number of players, so the approach doesn’t scale linearly. MultiGen maintains an explicit shared world state but doesn’t separate simulation from rendering. Agora-1 decouples the two and learns each as a separate function. One model evolves a shared, manipulable world state from player actions; a second renders consistent views of that state from independent viewpoints. The closest analogue is a modern game engine, only with both halves of the engine learned from data rather than hand-authored.

Learning through discovery

PROWL, released last week, gives world models a way to find their own failure modes and generate training data from them. None of these are products. They are the substrate for a class of interactive system that does not yet exist at scale: games that are generated rather than authored, robots that train in shared synthetic environments before they touch the real world, and foundation agents that grow up inside open-ended simulated worlds rather than on static datasets.

Learn more at RAAIS 2026!

Jeff Hawke, Odyssey’s co-founder and CTO, will go deep on this work at RAAIS 2026 in London on June 12.

Agora-1 can be played at agora.odyssey.ml. The Starchild-1 preview and technical report are live. Odyssey-2 is available via API at developer.odyssey.ml.

The bet behind world models is that the next leap in machine intelligence comes from interacting with a world, not from reading about one. Today, that world has sound and room for more than one.

This morning, Air Street Capital portfolio company Profluent announced a multi-program strategic partnership with Eli Lilly to develop AI-designed recombinases for genetic medicine. Profluent will receive an upfront payment, committed R&D funding, and is eligible for up to $2.25 billion in development and commercial milestones, plus tiered royalties on net sales.

This is the story of one of the hardest unsolved problems in genetic medicine kilobase-scale DNA editing - that finally has a credible path forward.

Kilobase-scale editing is the frontier

CRISPR was the breakthrough that taught us how to read and cut DNA at any address in the genome. It is a remarkable tool. It is also, in its dominant clinical incarnation, a typo corrector. Cas enzymes excel at breaking genes or, with engineering, making small changes to a single base or a short stretch. That solves a meaningful subset of genetic disease: the subset where one mutation, in one place, drives one phenotype.

The harder subset, and arguably the larger one, is genetic disease driven by heterogeneity: hundreds or thousands of different mutations across a patient population, scattered across the same gene. Cystic fibrosis is a textbook case. Many forms of inherited hearing loss, retinal dystrophy, and metabolic disease look the same. You cannot afford to develop a separate base-editing therapy per mutation. The economics never work.

The way through is to stop fixing typos and start replacing the whole paragraph: insert a healthy copy of the gene, in its correct genomic location, in one shot. Kilobase-scale, programmable, precise DNA insertion. This has been the holy grail of genetic medicine for as long as genetic medicine has existed.

It has also been more or less out of reach until now.

Subscribe now

Why recombinases and why they were stuck

Long before CRISPR, biologists knew about a class of enzymes called recombinases that do exactly the kind of large-scale DNA cutting and pasting that kilobase editing requires. Recombinases are precise, they are programmable in principle, and they have been used for decades as research tools.

CRISPR’s targeting is outsourced to a guide RNA: the enzyme stays the same; you change the guide; you go anywhere in the genome. Recombinases have no such modular guide. Their specificity is built into the protein itself, encoded in the three-dimensional shape of the enzyme. Retargeting a recombinase means redesigning the protein.

For most of the past few decades, redesigning a recombinase to hit a specific human genomic site with clinical-grade specificity was either impossible or so laborious it was not worth attempting. The field tried directed evolution. The field tried hand-crafted protein engineering. The results were narrow, slow, and difficult to generalise.

This is exactly the kind of problem that gets unstuck the moment you have a frontier model for proteins.

Gene editing is now an AI problem

Profluent’s is that protein design is a frontier AI problem, not a biology problem with AI bolted on. The company trains large frontier models on the world’s largest protein dataset, including the most comprehensive curated database of naturally occurring recombinases. It then conditions those models to generate novel enzymes for targets of interest: proteins that, in many cases, do not exist in nature and that would not have been found by searching nature for them.

The 2024 work that put Profluent on the map was the first public proof point of this thesis: AI-designed Cas proteins, built from scratch, that work. The lesson was that the same approach that worked for Cas should generalise to any class of designable protein where the data and the objective function are clear.

Recombinases are a near-ideal target for that thesis. There is enormous natural diversity to learn from. The substrates and target preferences of those natural recombinases can be matched to provide rich training signal. The goal is to design a protein that cuts and pastes a long stretch of DNA at a chosen genomic address with high specificity. This is a problem whose physics, data, and objective are all well set up for generative modelling.

If Profluent are right, kilobase-scale editing transitions from a discovery problem - sift through nature, hope you get lucky - into a design problem. You name the address; the model generates the editor.

Subscribe now

Why Lilly

The right partner for a platform play in genetic medicine is the company that is most aggressively building out the full clinical and commercial stack to take genetic medicines to patients. Lilly fits the description. Over the last few years they have stood up a dedicated genetic medicine center, acquired multiple in vivo gene and cell therapy companies, and signed a string of AI-native R&D collaborations. They are systematically assembling the components needed to industrialise genetic medicine - for rare disease today, and for a much larger set of common diseases as the toolkit matures.

A partnership of this shape, being multi-program, exclusive licensing on selected programs, $2.25B in milestones plus royalties, is what platform validation looks like.

What the future looks like

Profluent is building a programmable platform: name a genomic address, name the desired insert, get a designed editor whose properties are known in silico before anyone steps into a lab. Combine that with the in vivo delivery capabilities the field is rapidly maturing, and you have, for the first time, a credible path to therapies for diseases where the underlying genetics has always been the obstacle. These include heterogeneous monogenic disease, large-payload corrections, multi-gene insertions, and ultimately common diseases with structured genetic risk.

The next generation of gene editors will not be the ones we found in nature. They will be the ones we designed.

I am extremely proud to have been working closely with Ali and his team since day 1 of Profluent and making the company the largest position in the Air Street portfolio over the years. Today is a milestone for them and for the field - and a strong signal of where AI-designed biology is heading next. Congratulations to the entire Profluent team, and to the partners at Lilly stepping in at exactly the right moment.

When I started investing in 2013, deep learning was largely confined to research labs. Yet I was convinced back then that the most important technology companies of our generation will be built AI-first. This is because AI is a force multiplier for technological progress, and everything around us is ultimately a product of intelligence.

So in 2019, I founded Air Street to build a venture firm dedicated to this conviction.

Today, AI-first companies are emerging across every industry. We’re seeing the birth of highly capable long-horizon coding agents, AI-first vertical software products that produce human-grade work, design novel biological systems that outperform nature, control a wide range of embodied physical systems, and reshape - and test - our national security. The AI frontier has become the center of gravity of the technology industry, and by extension, every industry.

With Fund III, Air Street is now the largest solo GP venture firm in Europe. This structure enables high-conviction investing with a single decision-maker and significant capital to support the most ambitious teams building AI-first companies.

What our founders are building

Over the past decade, through my earliest angel investments and across Air Street’s funds, I’ve backed teams building across software, the physical world, science, and defense.

In software, Synthesia’s AI video platform now generates >$150M in annual recurring revenue with customers across more than 90% of the Fortune 100. Black Forest Labs‘ FLUX models have become the most widely used open visual intelligence models among developers and enterprises alike. Poolside has built a frontier AI lab operating in high-stakes enterprise and government environments.

In science, Profluent has created the first AI-designed CRISPR system (published in Nature) and demonstrated that scaling laws enable the design of ever more capable genome editors. Meanwhile, Enveda Biosciences is unlocking the chemistry of nature with AI to discover new medicines from natural products.

In the physical world, Wayve‘s AI driver operates zero-shot in over 500 cities around the world. Sereact deploys embodied AI robotics inside warehouses for leading industrial brands including BMW Group and Daimler Truck. And Lambda and Crusoe are powering the largest computing infrastructure buildout in modern history to train and deploy frontier AI.

And in defense and national security, Delian Alliance Industries is deploying autonomous defense systems to protect Europe and its allies at a time of critical need.

We have also seen several exits across Air Street funds, including Valence Discovery to Recursion Pharmaceuticals, Graphcore to SoftBank, Adept to Amazon, Allcyte to Exscientia (pre-IPO, subsequently Nasdaq-listed and merged with Recursion), Fern Labs to Poolside, and ClipDrop to Stability AI and later Jasper.

What Fund III enables

With Fund III, Air Street will lead early-stage rounds for AI-first companies in North America and Europe with checks of $500k to $15M. We will also make a small number of growth-stage investments with checks of up to $25M. Many of the teams will be well known to us, having built relationships over many years, through our network and events, and even prior investments in them.

Our focus remains on AI-first software, developer tools and infrastructure, techbio and science, and defense and security - the areas where AI is not simply improving the status quo but enabling entirely new kinds of products and companies to be built.

Building the AI community

From the beginning, Air Street has been about more than the companies we invest in.

The State of AI Report has grown into the most widely read and trusted annual analysis of progress in the field, now in ninth edition. RAAIS, our Research and Applied AI Summit, has brought together leading researchers and builders since 2017 - from the founders of DeepMind and the co-author of “Attention is All You Need” to the next generation of AI-first founders from Wayve, ElevenLabs, and Isomorphic Labs. The Air Street AI Network now connects nearly 3,000 vetted researchers, engineers, product leaders, and founders around the world from our global meetups in London, New York, San Francisco, Paris, Munich, and Zurich.

We will also remain committed to speaking out when the ecosystem needs it. Whether that is our work on spinout.fyi - where key Air Street reforms to the university spinout playbook were accepted by the UK Government - or our Air Street Press essays and policy memos, we believe that if something is worth saying, it’s worth saying publicly.

Our third epoch begins today

Fund III launches into what I believe is the most transformational period in technology of our generation. Capabilities that seemed like magic a decade ago are now real, usable, and creating enormous value. Air Street is built for exactly this moment, and I couldn’t be more energised to find and back the defining companies of this era.

I am deeply grateful for the trust and partnership of an extraordinary group of long-term partners including US university endowments, foundations, hospitals, and institutional investment platforms. Many of our existing investors significantly increased their commitments, and several are investing in a solo GP venture firm for the first time.

Air Street could not have made this progress without the hard work and ambition of our founders. It is a privilege to partner with you.

To the steadfast AI founder: come join our third epoch. We begin today.

tl;dr Air Street Capital, founded by Nathan Benaich, has raised $232M for Fund III - making it the largest solo GP venture fund in Europe. The fund invests in AI-first companies in North America and Europe across software, science, the physical world, and defense, with portfolio companies including Synthesia, Black Forest Labs, Sereact, Profluent, Delian Alliance Industries, and Poolside. Read more at www.airstreet.com and press.airstreet.com

In 2025, AI entered large-scale commercial deployment. Reasoning models made inference-time compute a new scaling axis, AI coding agents pulled automated software engineering into the present, and embodied AI alongside AI-accelerated scientific discovery emerged as major new battlegrounds. China’s open-weight ecosystem rapidly closed performance gaps with Western frontier models, even as the largest labs continued to widen their lead on capability-per-dollar. Scale remained the dominant force.

The economics followed. AI labs and public AI infrastructure companies added more net new revenue than the entire public SaaS sector. NVIDIA pushed beyond $4 trillion in market capitalization and catalyzed outsized returns across the AI semiconductor supply chain. Multi-gigawatt compute clusters moved from announcements to physical sites constrained by power availability, grid access, and geopolitics. AI infrastructure became industrial policy, with sovereign capital, export controls, and national champions shaping who can train, deploy, and profit from frontier systems. The US, in particular, moved forcefully to treat AI as a key competitive vector.

It is still early innings. Capabilities that would have seemed like magic a decade ago are now real and usable, and the opportunity ahead lies in deploying them widely and embedding them into the systems that matter. The next phase will be determined by who can convert frontier capability into durable, widespread use across economically meaningful tasks under real-world constraints of energy, economics, and politics.

In our 2025 year in review, we cover Air Street Capital, Air Street Press, the State of AI Report, and our community engagement.

Portfolio updates: AI investments across defense, healthcare, robotics, and fintech

In 2025, we made eight new investments across gaming, defense, healthcare, frontier AI and fintech, including Studio Atelico, Polar Mist, Delfa, Clove, and four companies yet to be announced. We also made two follow-on investments in techbio and defense, Profluent and Delian Alliance Industries, bringing Fund II to 18 active portfolio companies. During the year, we exited Fern Labs, a long-horizon agent company, to frontier AI company, poolside, via M&A.

Now turning to the portfolio highlights:

New investments:

Fern Labs: $3M Seed (solo investor). Fern Labs is the platform where networks of agents can create, test and iteratively improve software artifacts over long time horizons. The founding team previously built agent-driven AI systems at Palantir.

Polar Mist: Day 1 financing (co-lead investor). Polar Mist is building maritime autonomy for European defense. The company builds full-stack unmanned maritime defense, combining on-board jamming-free navigation with robust hardware platforms for long-range autonomous reconnaissance, logistics, and strike missions.

Delfa: $3.8M Seed (lead investor). Delfa is the AI-first operating system for clinical trials, starting with participant enrollment, where delays and cost are still driven by spreadsheets, phone calls, and duplicated site-by-site processes. Delfa’s agents are already deployed across 50+ live clinical trials trials.

Studio Atelico: $5M Seed (lead investor). Studio Atelico is an AI-first games studio. The founding team combines frontier ML and on-device optimization experience with award-winning game design, spanning Uber AI Labs, Meta, and Creative Assembly’s Total War franchise.

Clove: $14M Seed (co-investor). Clove is building an AI-native wealth management institution. The platform will pair regulated human advisors with an AI-first operating environment that absorbs repetitive and compliance-heavy work so advisors can serve more clients with consistent, transparent delivery.

Black Forest Labs: $300M Series B (co-investor). Black Forest Labs, founded by key inventors behind latent diffusion and Stable Diffusion, is translating research leadership into a category-defining visual intelligence company. The company has rapidly won Fortune 500 adoption and deep enterprise integrations as well as more than 400 million downloads of its open-source category-leading FLUX models.

Follow-on investments:

Sereact: €25M Series A (follow-on). Sereact is the frontier robotics research and deployment company, starting with warehouse automation. The platform supports picking and packing, quality checks, sorting, and inventory workflows.

Hedera Dx: €15M Series A (follow-on). Hedera Dx is scaling decentralized, hospital-run liquid biopsy testing using circulating tumor DNA, shifting advanced cancer diagnostics from a small number of centralized reference labs into routine clinical care across hospital networks.

Delian Alliance Industries: $14M Series A (co-lead investor). Delian builds vertically integrated, AI-first defense systems that autonomously sense and strike across land, air, and sea. Its full-stack portfolio spans GPS-denied navigation, electronic warfare, and autonomous effectors, built in Europe to defend Europe and its allies.

Profluent: $106M financing (follow-on). Profluent is a frontier AI company focused on writing biology. They’re scaling protein language models trained on the world’s largest private corpus of natural proteins (the Profluent Protein Atlas). The company has released the open source OpenCRISPR-1 AI-designed gene editor and entered into a handful of commercial deals spanning drug discovery and agriculture.

Exits:

Poolside (Fund I) acquired Fern Labs (Fund II) to deepen its push into long-running, reliable enterprise agents, bringing together agentic infrastructure and frontier models under one roof. Performance, reliability, and developer ergonomics are converging into a single stack that the winning frontier labs will own end-to-end. This deal was a special one for Air Street Capital, as we’re first investors in both poolside and Fern Labs.

Angel and pre-Air Street portfolio: AI infrastructure, enterprise software, and space

Across my angel and pre-Air Street portfolio, 2025 was a year where many AI-first companies scaled in earnest:

As the US pushed to rebuild domestic energy and AI capacity, Crusoe and Lambda both raised large rounds to meet demand for compute. Crusoe closed a $1.375B Series E at a $10B+ valuation to expand its vertically integrated AI factories, now underpinning OpenAI’s Stargate. Lambda raised a $480M Series D followed by a $1.5B+ Series E to deploy gigawatt-scale supercomputers for customers training and serving frontier models.

Elsewhere, Enveda raised a $150M Series D at a $1B valuation to advance its clinical pipeline built on AI-enabled natural compound discovery. Muon Space raised a $146M Series B to scale satellite manufacturing as demand grows for sovereign, vertically integrated space systems.

In enterprise software, Synthesia raised a $180M Series D as AI video became a default format for enterprise communication, now used by over 60% of the Fortune 100. ElevenLabs announced a $180M Series C as it closed in on $200M ARR for its AI-first audio products. PolyAI raised an $86M Series D as its voice agents handled tens of millions of high-stakes calls across healthcare, finance, and hospitality. incident.io raised a $62M Series B to build agents that actively resolve incidents with engineering teams. Motion raised $60M across Series B and C, scaling an agentic work platform used by over 10,000 businesses. Finally, Numerai raised a $30M Series C to continue running an AI-native hedge fund at scale.

Two companies were acquired in 2025: Niantic, creator of Pokémon GO, was acquired by Scopely for $3.5 billion, closing out one of the most successful stories in blending AI, games, and the physical world of all time. Ravelin, an AI-first fraud prevention platform for online commerce, was acquired by WorldPay.

Air Street Press

Air Street Press is the home for all of our analytical essays on AI research, industry playbooks as they take shape, policy memos, monthly State of AI newsletters, the annual State of AI Report, and our global event series.

In the last year, Air Street Press has clocked over half a million views, up almost 50%. Our most popular work included:

• 2025 is the year of open-endedness

• Sovereign AI is political branding

• The AI factory illusion

• AI progress, after 2025

• Is the EU AI Act actually useful?

• Can AI discover new science?

• The AI rollup mirage

State of AI Report 2025

In October, we released the eighth edition of the State of AI Report, the most widely read and trusted annual review of progress across AI research, industry, politics, and safety. The 2025 edition focused on the emergence of reasoning models, the economics of frontier systems, the resurgence of open-weight models, power as a binding constraint, and the acceleration of sovereign AI efforts among governments with the capital and political will to pursue them.

The report reached an even broader audience than in prior years and was a discussion springboard across leading AI labs, startups, governments, and investors. It was debated, critiqued, and pressure-tested in public and private, which is precisely its role.

To mark its release, we hosted State of AI launch events in San Francisco and New York, bringing together founders, researchers, and operators to discuss what the year’s progress actually implies for deployment, competition, and policy.

Those conversations continued across podcasts, panels, and closed-door sessions in the months that followed, informing how many in the field are thinking about what comes next. You can tune into podcasts I joined on TBPN, Matt Turck’s MAD podcast in New York, Turner Novak’s The Peel podcast, Daniel Bashir’s The Gradient podcast, and TechBio Talks with Chris Gibson (Recursion) in New York.

Community: AI researchers, founders, and operators

In 2025, we continued to spend time where we learn best: in rooms with people actually building and deploying AI systems. Through Air Street AI meetups and the ninth edition of the Research and Applied AI Summit (RAAIS), we brought together hundreds of researchers, founders, and operators to compare learnings on what is working, what is breaking, and what is changing faster than expected.

We keep RAAIS and Air Street AI meetups intentionally small and highly curated, with an emphasis on open discussion. Conversations in 2025 ranged from AI-first biology and coding agents to defense, open models, and the realities of scaling systems under real-world constraints. The goal is simple: help people learn faster from one another and leave with ideas they can apply immediately.

RAAIS remains a highlight of the year, which you can watch on our YouTube channel and read the takeaways on Air Street Press. All proceeds support the RAAIS Foundation, which funds open educational resources and works to broaden participation in advanced AI research. We’ll return on 12 June 2026 for the tenth edition and continue to host events across Europe and North America throughout the year. Join us in Munich (17 Feb), Zurich (19 Feb), Paris (11 March), San Francisco (29 April), and London (12 June).

Here’s to a big 2026!

If 2023 and 2024 were about discovering and pushing what AI could do, 2025 was about learning where it actually works in the real world at scale. The next chapter will be defined less by frontier breakthroughs and more by diffusion: who can make AI dependable, affordable, and embedded in the systems that matter. That is where we continue to focus, and where we are most excited to build. We can’t wait to see what the community achieves in the coming year!

As ever, drop me a reply on nathan@airstreet.com if you’re building in AI.

To the AI community, the names Robin Rombach, Andreas Blattmann, and Patrick Esser sound very familiar. In the summer of 2022, their work on Latent Diffusion and Stable Diffusion bent the arc of visual AI by ushering photorealistic image generation onto the internet. The world went nuts, hundreds of thousands of developers built with the models, and Twitter feeds could speak (or show) nothing else. Indeed, this proceeded the low key research preview of ChatGPT, which celebrated its third birthday only yesterday.

So for the trio that ignited a revolution in generative media, founding Black Forest Labs (BFL) was the natural next step. And their focus was to translate research leadership into a category‑leading visual intelligence company built for the next decade of multimodal computing.

Today, I’m excited to unveil Air Street’s investment in Black Forest Labs as it announces a landmark funding milestone: a $300M Series B, following a previously unannounced Series A. Together, these rounds represent a big step in scaling the company’s momentum, with Black Forest Labs now capitalised with half a billion dollars and trusted by leading Fortune 500 enterprises.

From breakthroughs to exceptional execution

In under 14 months, Black Forest Labs has delivered one of the most impressive execution runs we have seen in generative AI. The company has gone from zero revenue to significant scale with Fortune 500 adoption, deep enterprise integrations, and use cases ranging from e-commerce to robotics simulation. More than 400 million downloads of their open-source FLUX models make them the most widely deployed visual-generation systems in the world. FLUX has become shorthand for quality, controllability, and reliability.

The company exhibits a set of attributes that are very rare to find and are the result of sustained, disciplined ingenuity and velocity across research, product, and go‑to‑market:

• Frontier AI, open by default. Black Forest Labs defined what open frontier visual AI should look like: high‑quality, open base models with transparent research lineage and enterprise‑grade deployment options that meet the standards of global companies.

• A new standard for controllable image editing. FLUX.1 Kontext and FLUX.2 transformed image generation from a probabilistic slot machine into a precise creative instrument, effectively creating a new standard for professional image editing workflows.

• Compounding enterprise traction. From social platforms to global retailers, Black Forest Labs systems are already embedded in high‑value, high‑volume production workflows that demand reliability, governance, and predictable performance.

• A world‑class team scaling with purpose. In under a year, the team has grown from 15 to more than 40 researchers, engineers, and operators from NVIDIA, Meta, Databricks, Google, Spotify, AWS and more, forming one of the most concentrated talent clusters in visual AI.

• An enormous market. It would not be a stretch to say that billions of images are uploaded to the internet every day. And with visual AI progressing at such a pace, a large proportion of these and new content will switch to being generative from the outset.

Watch co-founder Andreas Blattmann’s talk at our 10th annual Research and Applied AI Summit in London, 13 June 2025:

The road ahead: multimodal intelligence and beyond

The roadmap reflects the Black Forest Labs’ intent to define - not simply participate in - the next frontier of visual AI. FLUX.2 launched just last week and offers impressive precision, efficiency, control and extreme realism at a fraction of the cost of systems like Google’s Nano Banana Pro. Try it on the playground!

Next up will be more powerful multimodal models that offer market-leading infrastructure to underpin how billions of people and systems create, reason, simulate, and communicate.

This shift is already visible in early customer deployments: models are moving from generating images to supporting full creative workflows, interactive video, and multimodal understanding. As the boundary between perception and generation collapses, the winners will be the companies that combine frontier research with the reliability and depth required by the world’s largest enterprises.

Black Forest Labs is that company.

Into the New

Black Forest Labs is is building the visual foundation of the next decade of AI - open, controllable, multimodal, and trusted in the most demanding settings. We’re proud to be working with Robin, Andreas, Patrick, and the entire team as they continue to define what visual intelligence can be.

🌲 Into the New 🌲

For more than three centuries, wealth in the United Kingdom has been shaped by institutions like Coutts, founded in 1692 and long associated with a model of personalised service reserved for a narrow segment of society. That heritage has defined expectations of advice as something scarce, relationship driven and costly to deliver. Today that world looks very different. The mass affluent have grown in number and complexity, accumulating meaningful assets while facing a financial landscape defined by rising volatility, shifting pensions, tightening tax regimes and the largest generational transfer of wealth in modern UK history. At the same time, the institutions that should support them have not kept pace. Advisors are scarce, legacy systems are slow and the burden of navigating important long term decisions has fallen back onto individuals.

It is against this backdrop that Air Street Capital is investing in Clove’s $14M first financing round, backing a new kind of wealth institution built for this generation rather than the last. This investment reinforces our long-term thesis that AI-native financial institutions will reshape how wealth is planned, managed and delivered.

Clove’s founders, Christian Owens and Alex Loizou, see this gap not as an inevitability but as the result of infrastructure that was never built for the modern consumer. With Clove, they are creating a new kind of financial institution, one designed from the ground up for people who want trustworthy guidance but have been priced out or ignored by traditional services. Their platform brings together regulated human advisors with an AI-first environment that handles the repetitive and compliance heavy processes which dominate advisory work today. By removing friction and expanding advisor capacity, Clove can deliver high quality personalised guidance at a scale that has not been possible before.

Rebuilding the wealth stack

The scale of the advice gap in the United Kingdom captures both the urgency and persistence of this problem. Thirteen million households hold between £50k and £5M in investable assets, yet nearly three quarters of them receive no professional guidance. Many are left to navigate pensions, taxes and long term decisions alone, despite the growing complexity of the financial landscape.

This lack of support is reflected in the way people use even the simplest tax efficient investment tools such as the ISA. Only around 40% of adults hold an ISA and fewer than 10% of ISA holders use their full £20,000 personal allowance each year. This illustrates a broader pattern in which millions of people with meaningful savings struggle to make informed choices. DIY investing platforms were designed for trading rather than long term planning and regulatory efforts aimed at transparency cannot compensate for the operational limits of institutions that still rely on outdated processes.

Clove is stepping into this gap with the goal of widening access to high quality financial guidance. Indeed, only around 8.6% of UK adults received regulated financial advice in the twelve months to May 2024, and half of advisers have stopped serving less wealthy clients. The Financial Conduct Authority estimates that 7M adults with at least £10K in cash savings may be missing out on the benefits of long term investing. Clove aims to close this structural gap by building a modern institution able to support a far broader population with consistency, trust and transparency.

Our long-term partnership

We have known Alex for more than a decade and backed him before. He and Christian bring the experience of building and scaling complex fintech and marketplace platforms. Christian founded Paddle and scaled it into a global payments infrastructure platform serving software companies around the world. Alex founded Trouva which created a novel consumer marketplace for independent retailers in Europe. Their shared background in regulated operations, financial infrastructure and global distribution positions them well for the challenge of rebuilding the wealth industry. Both founders have also experienced the gaps in the current advisory market personally which gives them strong conviction and a clear understanding of consumer needs.

A new standard for wealth advice

Air Street Capital invests in Clove as the first AI-native wealth institution designed for the mass affluent. This investment reflects our belief that category-defining companies in financial services will be AI-native from day one. We believe Clove represents a step change in how financial guidance will be delivered over the next decade. The company combines the humanity and regulatory discipline of professional advisors with the scale, consistency and analytical power of AI. This combination has the potential to expand access to high quality advice for millions of households who have been underserved for years.

As AI transforms financial planning globally, Air Street Capital is committed to backing the institutions that will define this new category. We are excited to support Christian, Alex and the Clove team as they build a new institution for the mass affluent and set a new standard for modern wealth management.

Subscribe now

I first reached out to Ali Madani in 2021 via a Twitter DM after featuring his team’s breakthrough work on AI‑generated proteins in that year’s State of AI Report. It caught my attention as one of the first real proofs that generative models could produce functional biological machinery that outperformed naturally occurring variants. The Salesforce Research preprint, later published in Nature Biotechnology, showed that large language models could generate de novo proteins that functioned as well as, or better than, those found in nature. It was the clearest evidence yet that biology could be written with intent rather than discovered by chance. A few months later, Ali founded Profluent, and I wrote one of the largest first checks in Air Street’s history.

Now, that early idea has become a defining company at the frontier of AI-first biology. Profluent has raised $106M led by Bezos Expeditions and Altimeter, with continued support from Air Street, Spark, and Insight. The company is now the largest position in Air Street’s second fund. This new capital, which brings the total raised to $150M, accelerates the company’s path toward scaling frontier protein models and, ultimately, delivering the first AI-designed therapeutic to a human patient.

The shift from serendipity to design

For decades, biological progress relied on chance discoveries: CRISPR stumbled upon in bacteria, antibodies evolved through slow mutagenesis, enzymes refined through years of directed evolution. In 2019, I argued that the next generation of biotech companies would break from this pattern and move toward deliberate, AI-driven design. Profluent is the strongest realisation of that shift.

In just three years, the company has built the world’s largest private corpus of natural proteins, the Profluent Protein Atlas, and demonstrated - clearly and repeatedly - that scaling laws apply to biological design. ProGen evolved into ProGen3, a frontier-scale model capable of generating functional proteins across diverse families. Larger models predict function better and produced proteins that folded, bound and acted in cells in ways nature never explored.

Breakthroughs that changed the field

The moment that crystallised Profluent’s leadership was OpenCRISPR‑1, the first AI-designed genome editor. Released openly, it was downloaded tens of thousands of times within a day, with more than a hundred licence requests from pharma, agriculture, biotech, and academic labs. For the first time, a large model had generated a genome-editing molecule from scratch.

Profluent has since expanded the frontier with E1, the first retrieval-augmented model for protein engineering - bringing techniques from natural language systems directly into biological design. And with Protein2PAM, the team made gene editors programmable with high precision by designing the exact DNA motifs they can recognise. This unlocks edits at genomic sites previously out of reach and extends the company’s capabilities into antibodies, enzymes and other high‑value modalities.

Profluent’s scientific momentum has translated into commercial traction across multiple industries. A multi‑year collaboration with Corteva Agriscience is applying frontier gene-editing tools to next‑generation crops. A strategic partnership with Revvity combines Profluent’s AI-engineered enzymes with the Pin‑point™ base‑editing platform to produce more precise adenine deaminases. And with Integrated DNA Technologies, Profluent is co‑designing enzymes that move seamlessly from in silico creation to ready‑to‑use reagents for researchers.

What this capital unlocks

This $106M round enables Profluent to scale its foundation models even further, expand its automated wet‑lab capacity, and strengthen its commercial footprint across three major verticals. Most importantly, it brings the company closer to an extraordinary milestone: dosing the first human patient with an AI‑designed therapeutic.

My conviction in Profluent has only deepened over the years of our collaboration and work to get to this inflection point. Ali and the team have shown technical originality, scientific rigour, and execution speed at a level rarely seen. Profluent is defining the shift from reading biology to writing it.

I’m proud to continue supporting them as they push programmable biology into its next, most consequential chapter.

When I first met Ash in late 2024 through a mutual friend at Palantir, Fern Labs was still operating under a codename. Despite the early state of the company, one thing was obvious: Ash had the rare combination of technical depth, velocity, and clarity of product intuition that you only see in the very best AI founders. His colleagues at Palantir described him as someone with exceptional taste, the ability to lock onto the right problems before others see them, and the person who repeatedly pushed the organisation toward generative AI long before it was fashionable. This is exactly the founder phenotype that we look for at Air Street: technical, pragmatic, and customer value-centric.

With Fern, Ash was pushing toward something more ambitious and more precise than yet another agent framework. His goal: to solve long-running, reliable agents by creating an agentic operating system that fused structured reasoning, tool use, and traceability. It echoed many of the ideas I wrote about in Open-endedness is all we’ll need, which argued that the next stage of truly usable AI agents will come from systems that continually generate novelty, self-improve, and orchestrate long-horizon tasks in unpredictable environments. Ash had independently arrived at similar conclusions long before we ever compared notes, particularly the importance of systems that can adapt, iterate, and expand their own capabilities over time.

Over a quick succession of calls and async docs, Ash explained how Fern’s architecture was built around this exact intuition: long-running loops of generation, critique, repair, memory, and tool use that allow agents to improve across episodes rather than simply complete a single task. Indeed, Fern was an early glimpse of the systems that will eventually handle complex, multi-stage real‑world workflows. Succeeding at this mission would be of critical importance to the success of agents in the enterprise.

Without hesitation, Air Street led the entire $3M financing to back Ash before the company had a website, a deck, or a team. Shortly thereafter, Alex and Taylor - two equally talented Palantirians - jumped into Fern and we were off to the races. Here’s what happened next.

From short demos to 10-hour autonomous builds

One of the cleanest ways to understand Fern’s progress is through time horizons of autonomous agent work. Around the time METR published excellent work on measuring how long current agents can run before falling apart, even the best models like Sonnet 3.7 only reliably handled tasks under an hour - a far cry from real-world workflows. Meanwhile, Fern was already pushing well beyond those limits.

In the early months, the team focused on short autonomous builds that were concrete enough to stress-test the system end to end. Fern would spin up several agents with their own virtual machines, access to tools and a product requirement document for what the user wanted Fern to create. Fern agents built products like a remote interviewing tool in roughly three hours of autonomous work, a Slack-style workspace clone with real functionality, and a functional video calling site with screen sharing, recording and no noticeable latency on live cross-continental video calls. Taken together, these showed that Fern could wire together non-trivial features across frontend, backend, and data.

By the time Ash joined us at the Air Street NYC AI meetup in mid-April 2025, Fern was routinely running agents for several hours at a time, up to 10 hours - a feat that the field had not yet demonstrated publicly. In that talk, he showed something important: when you let agents run longer, the quality of the resulting software build compounds. With more time, the same Fern agents produced more polished UIs, discovered and implemented advanced features like search, authentication, and richer workflows that you would normally only expect from a human team iterating over multiple sessions. At a moment when the best public evaluations suggested that agents struggled beyond a few hours of equivalent human work, Fern was quietly demonstrating 5x that horizon in the wild.

The next chapter

In tandem with Fern accelerating its revenue build via a forward deployed motion with large enterprises, frontier labs were expanding their agentic capabilities. It was becoming clear that winning the largest prize in AI - agents that write, maintain and improve all software - requires compounding compute, models, tools, and workflows into a unified development environment. Put another way, a frontier AI company would need Fern, just as Fern would need a frontier AI company.

Meanwhile, poolside - the frontier AI company tackling the most high-stakes deployments - had been pursuing exactly this vision. I’ve been involved with poolside since before its creation and as one of the first investors with Air Street’s first fund. I have seen firsthand how Eiso and Jason built the company from a set of clearly articulated beliefs - including that mastering software development is a viable path to AGI - into a real contender at the frontier.

Following an introduction I’d made between Eiso and Ash, the two companies began exploring what a go-to-market partnership could unlock in each other’s roadmaps. It rapidly became clear that joining forces would help both companies reach their shared roadmap far faster.

What poolside and Fern are building together

Poolside is building one of the strongest full-stack AI companies in the world by uniting the energy, compute, models, and infrastructure needed to run multi-agent systems for complex enterprise workflows under one roof. The Fern team brings a deeply opinionated agentic core - built from first principles and stress-tested on real workloads - into a company with scale, distribution, and compute firepower. Fern’s architecture was built for agent specialization and coordinated long-horizon work - the exact capabilities poolside can scale into a full production environment.

Together, they can move faster on:

• Agentic infrastructure that actually scales, not just demos well.

• Enterprise-grade traceability and safety primitives, which Fern built from the ground up.

• Developer tooling that makes reasoning systems predictable, debuggable, and production-ready.

• Integrated model and agent loops, which poolside is uniquely positioned to own end-to-end.

This is the direction the market has been moving toward: a world where model performance, agent reliability, and developer ergonomics form a single continuum. Here, the winners will own that entire stack - a belief I have held for many years.

A word about the AI community Air Street is building

One of the privileges of working in AI for over a decade is seeing how networks of startups, bigger companies, and researchers compound. The opportunity to work with Ash on Fern existed because my network surfaced an exceptional founder when he took the risk to build something new. Poolside exists because that same community has been pushing the frontier of AI capabilities for years.

When these networks work well, moments like this emerge organically: founders meet each other, visions align, and the right strategic move becomes the obvious one. Indeed, the best outcomes don’t happen by accident - they emerge when the right people, the right ideas, and the right timing align. This is one of those moments.

I’m proud to have backed both entrepreneurs early. I’m even more proud to have helped bring them together.

And I’m excited for what they will build next! 👀

Understanding how protein sequence encodes structure and function remains one of the central challenges in the life sciences. Yet most protein language models still treat each sequence as an isolated datapoint. This forces the entire burden of evolutionary context into model parameters, which leads to blind spots in underrepresented families and amplifies the biases of sequence databases. Profluent’s (an Air Street portfolio company) new E1 family demonstrates that this constraint is no longer necessary. Retrieval augmentation, a technique that transformed natural language processing, is now beginning to reshape protein modeling by allowing models to incorporate evolutionary information at the moment of inference rather than storing it all in weights.

E1 is built around this shift. Its design embraces the idea that proteins are best understood in the context of their evolutionary families. Functional motifs, co-evolving residues, and structural constraints emerge most clearly when a sequence is compared to its close relatives. By enabling the model to draw on homologous sequences during training and inference, E1 brings this relational structure directly into computation.

How E1 works

The core architectural idea behind E1 is block-causal multi-sequence attention.

Instead of examining a single protein in isolation, the model receives a query sequence preceded by a set of unaligned homologs. It alternates between layers that attend within each sequence and layers that allow residues to attend across sequences. Because these homologs are unaligned, E1 must implicitly learn how positions correspond across evolutionary space. This mirrors the flexibility that made retrieval so powerful in NLP, where models learn to use external context without relying on rigid structure.

Training further reinforces this capability. Profluent trained E1 on 4 trillion tokens drawn from its proprietary Protein Atlas and UniRef, gradually increasing both the length of multi-sequence instances and the number of homologs included. The result is a family of models at 150M, 300M, and 600M parameters, all released under a permissive license for research and commercial use.

Subscribe now

E1 sets the new open-source standard

To understand how this works in practice, consider a protein engineer developing next-generation CRISPR genome editors. Profluent’s earlier model, OpenCRISPR1, showed that language models can learn meaningful relationships among thousands of CRISPR-associated nucleases with minimal supervision. Now imagine an engineer seeking to design a Cas effector variant with improved specificity for a therapeutic context. With E1, they could provide the wild-type nuclease sequence, retrieve its closest evolutionary neighbours from public databases, and ask the model to score how candidate mutations might affect activity or off-target behavior. Because E1 processes the nuclease in the company of its evolutionary relatives, it can highlight conserved positions that tolerate little variation, identify residues whose variation correlates with altered PAM preferences, and surface rare substitutions found only in distant homologs that may expand targeting range.

This kind of enriched reasoning is reflected in benchmark performance. On ProteinGym, E1 establishes a new high-water mark for open-source zero-shot mutation effect prediction. At equivalent parameter scales, all three E1 variants outperform both ESM-2 and ESM C. When homologs are included at inference time, E1 surpasses PoET and MSA Pairformer, reaching an average Spearman correlation of 0.477 and an NDCG at rank ten of 0.788 for the 600M model. The advantages are clearest in protein families with limited multiple sequence alignment depth, where single-sequence models often fail and retrieval provides critical additional signal.

Structural reasoning follows the same trend. Using the Categorical Jacobian method to estimate residue-residue contacts, E1 outperforms the ESM-2 family across every scale. On CAMEO, a continuous blinded benchmark for protein structure prediction, the 600M variant achieves a long-range precision at L of 0.512 in single-sequence mode and 0.541 with retrieval. Examples in the technical report show how retrieval sharpens contact predictions in regions where single-sequence models misinterpret noise as structure.

Taken together, E1 treats evolutionary relationships not as static information distilled during pretraining but as an active force shaping inference. This allows a single model to specialise dynamically for different families without fine-tuning, avoids catastrophic forgetting, and reveals clear scaling behaviour across the released model sizes.

Advancing open protein engineering science

The open release of E1 materially shifts what is possible for research groups and startups working in protein engineering. It offers state-of-the-art performance while remaining small enough to run locally, lowering the computational barrier that often limits exploratory design work. Because the model does not require aligned MSAs, it also simplifies the early phases of design when sequences may be diverse or poorly characterised.

More broadly, E1 shows how ideas first proven in natural language modeling are now regularly diffusing into protein engineering. Retrieval-augmented inference revolutionised document and knowledge retrieval by allowing models to draw on external context at runtime rather than relying entirely on what they had memorised. Profluent has now demonstrated that the same principle applies in biology. By giving a protein model the ability to pull in relevant evolutionary context on demand, E1 represents an important step toward biological foundation models that can flexibly integrate many sources of information.

This cross pollination between AI and biology, where architectural innovations migrate from language to sequence analysis, is likely to guide the next several years of progress. Retrieval may prove to be the missing connective tissue between today’s sequence-only models and future systems that reason jointly over evolutionary, structural, and experimental evidence.

Read the technical report

For PARIMA, the approval validates years of research and development, as well as industrial scaling work. The company has built one of the largest cultivated food facilities in the world, developed novel cell lines, as well as bioprocess engineering techniques and infrastructure for scalable production. Its category-leading regulatory portfolio now spans nine active filings across major jurisdictions. Moreover, PARIMA strengthened its supply chain by acquiring a key cell line supplier in the cultivated food ecosystem, Vital Meat, to further consolidate its position as a global leader in next-generation proteins.

But make no mistake: Singapore’s approval is no accident. Since 2019, the city-state has pursued a coherent national strategy for alternative proteins within its broader “30 by 30” food security plan aimed at producing 30% of the nation’s nutritional needs locally by 2030. This includes a predictable, science-led regulatory process and targeted incentives for industrial deployment. The result is a functioning market where new food technologies don’t just pilot. They sell.

Europe, meanwhile, continues to incubate world-class science without the political will to deploy, buy or scale it. The EU’s novel foods pathway has become a regulatory maze that would challenge the labyrinths of Greek and Minoan civilization in its complexity. According to Nature Food, the average time for approval of a novel food application in the EU now exceeds 35 months - a delay that effectively sidelines startups and deters investment. Worse yet, it rewards incumbents and retards innovation. PARIMA’s approval in Singapore is not a failure of European science. It is a failure of European policy.

Cultivated food is now entering its industrial era. The question is not whether the technology works - it does - but where it will scale. If Europe wants to lead in sustainable food production, it must match scientific excellence with regulatory competence. Otherwise, the continent will remain a laboratory for breakthroughs that commercialize elsewhere. And if Europe is content to let others commercialize its discoveries, it should be honest about that reality.

Read more from Parima

The story of cultivated meat has so far been one of scientific triumph and industrial constraint. Several startups have proven that animal proteins can be grown from cells rather than animals, yet few have crossed the chasm from lab bench to factory floor. The bottleneck is less about taste and technology and is more about scale and regulations.

This month, our portfolio company Gourmey acquired Vital Meat to form PARIMA, a new European champion for cultivated proteins and foods. The merger brings together two complementary pioneers: Gourmey, known for its work in high-value duck and poultry cell lines, and Vital Meat, focused on scalable cell-based chicken production. Together, they now control the critical layers of the cultivated stack, from cell biology to bioprocess design to regulatory readiness, while moving beyond scaled R&D into genuine food manufacturing. Its Paris facility, one of the largest of its kind worldwide, anchors a new phase of industrialization: large-volume production, supply chain integration, and cost parity with high-value meats.

What’s behind the announcement

PARIMA combines Gourmey’s full-stack industrial platform, including premium cultivated duck validated by Michelin-starred chefs and independently verified production costs below €7/kg, with Vital Meat’s Nantes infrastructure, namely its 2,000‑liter bioreactor capacity and advanced avian cell lines developed from ~25 years of Groupe Grimaud’s avian cell research. The combined platform is supported by 15+ patent families and 70+ patent applications, and a regulatory portfolio of nine active filings across major regions. PARIMA’s near-term goal is to become the first cultivated food company positioned for regulatory approval of two species (duck and chicken), with applications that span premium dining to large-scale B2B ingredients.

A milestone for Europe’s bioeconomy

Just as the AI industry is integrating vertically around compute and data, cultivated food is consolidating around production infrastructure and process efficiency. The winners will not be the first to demonstrate a prototype, but the first to make a tonne.

At Air Street, we’ve backed Gourmey since its early days because we believed that cultivated food would follow the same curve as other deep technology sectors: a burst of discovery, a period of consolidation, and then a race to scale. PARIMA embodies that transition: from a scientific breakthrough to an industrial one.

While you’re here: another milestone with Mission Barns hitting retail!

Alongside PARIMA’s formation, another Air Street portfolio company reached a landmark: Mission Barns has begun selling cultivated pork products to consumers in the U.S. Following an FDA “no questions” letter for its cultivated pork fat in March 2025 and USDA clearance in July 2025, Mission Barns launched with Bay Area restaurant partner Fiorella and is executing targeted retail releases. Notably, its cultivated pork meatballs are making a supermarket debut at Berkeley Bowl in the East Bay, one of the first times cultivated meat has been sold at a U.S. grocer.

When Air Street portfolio company poolside launched in 2023, it set out to build artificial general intelligence by mastering software development in the most demanding environments. Software, the company argued, requires reasoning, planning, and an understanding of the world. The same ingredients underpin human cognition. From the outset, poolside’s mission has been to translate those principles into systems that operate reliably and at scale, supported by infrastructure built to last.

Now that conviction takes physical form. Poolside has unveiled Project Horizon, a 2GW AI campus in West Texas designed from the ground up for multi‑agent systems. Through a partnership with CoreWeave, which will anchor and operate the first 250MW phase, poolside is combining compute, power, and intelligence under one roof to create a vertically integrated foundation for the agentic enterprise.

From data centers to AI factories: a purpose-built cloud for AI agents

As we wrote in the State of AI Report 2025, power has become the new bottleneck for AI progress as large labs compete for resources to train and serve the insatiable demand for their frontier models. Multi-GW clusters have moved from slideware to site plans, and grid constraints are beginning to shape both roadmaps and margins. What’s more, AI infrastructure has become a lever of national strategy. This means that the ability to generate, store, and apply compute power independently is a question of technological sovereignty.

Project Horizon is poolside’s answer to these challenges. It is a ground-up data center development designed to give the company long-term control over power, cost, and scale through vertical integration. The campus will be developed in eight phases of 250MW each, ensuring scalable, modular growth aligned with advances in compute demand and silicon efficiency. In doing so, poolside can commit to providing long-term access to energy, compute, and intelligence to serve the most demanding of customers.

Located in West Texas, the site sits near a major natural gas hub with access to existing water, processing, and fiber infrastructure. It occupies part of the Longfellow Ranch, whose owners, the Mitchell family, have supported Horizon’s development across their expansive property. Dual long-haul fiber routes provide high-bandwidth, low-latency connections to major network hubs, while the behind‑the‑meter design allows poolside to directly manage its power generation. The campus is engineered for modular expansion, built with redundancy at every layer of the power stack. Generation will come from efficient aero‑derivative turbines meeting strict emissions standards, supplemented by grid‑interconnect redundancy and on‑site battery storage to improve reliability and load balancing.

A milestone for poolside

For poolside, Horizon is an infrastructure milestone and a major expansion into controlling the inputs that determine who reaches the frontier. In an industry where access to energy and clusters now defines the limits of ingenuity, Horizon gives poolside direct control over the two scarcest resources in AI: power and compute. Owning this foundation means the company can train and serve frontier‑scale models on its own timeline, protect customers from the volatility of cloud pricing and availability, and invest in sustained capability growth rather than short‑term capacity.

This level of vertical integration - from model and agent development through to power generation and deployment - gives poolside a durable advantage as AI systems become indispensable to enterprise and national infrastructure alike. Project Horizon turns energy and compute from external constraints into internal levers of scale, marking poolside’s growth from an AI model builder into an industrial company of intelligence.

AI is the ultimate force multiplier on technological progress in our digital, data-driven world. Our mission at Air Street Capital has always been to back the most ambitious teams building breakthrough AI products that would have seemed like magic when I started investing over a decade ago. Key to these inflection points is the infrastructure that enables them: NVIDIA computing systems. I have long argued that NVIDIA is the defining company of the AI era, powering breakthroughs in science, industry, and national strategy. This is why I’m excited to share that we are deepening that story together: Air Street Capital is partnering with NVIDIA as part of its new £2B commitment to the UK AI ecosystem.

NVIDIA announced this week that it will invest in and support UK-based AI startups with a mix of capital, compute, and technical expertise. This builds on its role as the critical enabler of the global AI industry, which we’ve showcased in depth since 2018 in our State of AI Reports, as the demand for GPUs reaches new heights from researchers, startups, enterprises, and governments alike.

This partnership reflects the conviction that Air Street Capital and NVIDIA both share: the UK has what it takes to produce generational AI-first companies, provided founders can access world-class compute infrastructure and communities. In my 10+ year career investing in AI, I’ve backed many leading UK AI companies including ElevenLabs, Synthesia, Wayve, PolyAI, V7 Labs, Fern Labs, and Thought Machine. Alongside these investments, I’ve built communities like London.AI and the Research and Applied AI Summit, where the best in AI come together to share learnings and ideas to advance research and products.

Now, with the combined backing of NVIDIA and Air Street Capital, the UK AI ecosystem is primed to enter a new phase of growth and global impact.

Read more

That’s why we’re leading the $3.8M Seed round for Delfa. The team is building an AI-native operating system for clinical trials, starting with patient enrolment, the most broken part of the process.

The bane of spreadsheets and phone tag

There are 22,000+ drug candidates in the global pipeline but just ~5,000 new trials start each year. More than a third of the cost and delay comes down to a single operational challenge: finding, qualifying, and enrolling patients.

It’s still a manual funnel. Sites use Excel trackers. Recruiters cold-call lists, answer the same questions, collect medical histories by phone, and schedule visits by email. Every site solves the same problem in isolation.

Delfa solves recruitment with AI agents

Delfa replaces the human bottleneck with AI agents that fulfill the role of expert recruiters. They are trained with clinical trial documentation and guardrailed accordingly. Every patient interaction is logged, reviewed, and scored across 20+ QA metrics. Physicians audit the outputs. Alerts fire in real time when something looks off.

The AI agents can speak with patients in any language by phone, SMS, or chat; prescreen for medical eligibility; answer questions; collect histories; book visits; and write directly into a site’s CRM or CTMS.

In just a few months, Delfa’s solution is live across 50+ trials and multiple indications. Delfa works with top site networks like Flourish, Elevate, and Re:Cognition. Importantly, these site networks had the option of continuing with large recruitment teams and legacy tools, but chose Delfa instead. In one migraine study, a single site exceeded its enrollment target by 25%, saved 20 recruiter hours, and achieved an 11% lead-to-enrolled rate. In a vaccine trial, Delfa’s agent matched human conversion rates.

Why we invested

We’ve spent years backing the entrepreneurs building the future of drug discovery, including Valence Discovery (acq. by Recursion), Allcyte (acq. by Exscientia), and most recently, Profluent. That work gave us a front-row seat to how promising science gets bottlenecked in the clinic.

Delfa is the first team we’ve seen turn that bottleneck into software infrastructure. They’re translating the potential of AI agents into something that helps real patients, real sites, and real sponsors. But recruitment is just the beginning. Delfa is building a full-stack OS for clinical trials: from smarter site selection to automated study start-up and long-term patient retention. It’s a bet on turning clinical ops from artisanal workflows into scalable software.

We’re hiring!

AI agents are finally good enough to take on repetitive, high-volume, high-compliance workflows in healthcare. Delfa is one of the few teams translating that capability into measurable ROI, and doing it in a wedge that touches every new drug.

CEO Tom Farrand has built and deployed applied AI across J&J, Medtronic, and Exscientia. CTO Gustav Bredell trained at ETH Zurich, built early foundation models at Roche, and has deep experience in medical AI systems.

Come join the growing team in London and SF!

Read more in Upstarts Media

You can also read more from in who has the launch story:

Upstarts Media

The Startup Speeding Up Clinical Trials With Voice AI

The Upshot…

Read more

10 months ago · 6 likes · Alex Konrad

We’re excited to announce Air Street Capital’s lead investment in Studio Atelico’s $5M Seed round, alongside friends including Chris Ré (Stanford), Thomas Wolf (Hugging Face), and Alex Ratner (Snorkel), to back a team on a mission to redefine the role of generative AI in games.

Founded by Piero Molino, Paul Szerlip, JP Chen, and Ennio De Nucci, Studio Atelico brings together deep experience in AI model development, on-device optimisation, and award-winning game design. The founders met through years of collaboration at Geometric Intelligence (acquired by Uber) and forming Uber AI Labs together, where they built state-of-the-art ML systems including Ludwig. After Uber, Piero spent some time with Chris Ré (which is how I met him!) and then went on to found AI startup Predibase, which was acquired by NYSE-listed Rubrik in June 2025. Meanwhile, JP went to Instagram/Meta to work on large-scale on-device inference and generative AI. Over in London, Ennio led game design for Creative Assembly’s Total War series, including the Warhammer trilogy and Three Kingdoms, which sold millions of copies worldwide.

Why now?

The next era of games will be defined by generative AI. Not as a bolt-on, but as a core game mechanic. Yet today’s AI-powered games face steep hurdles: API-based inference costs eat into margins, latency interrupts immersion, and developers have little control over model behaviour.

Studio Atelico’s answer is an on-device AI engine that gives developers and itself as a studio full control. The engine powers lifelike NPCs with rich dialogue, planning, and emergent behaviours, all without server latency or per-interaction costs. It integrates directly into major game engines like Unreal, Unity, and Godot. The result: richer worlds, deeper player engagement, and the economics to scale.

We’re especially excited by Atelico’s approach to build and launch its own games on top of this technology. Gaming is a hits-driven business, and the value of a unique capability is best realised when the team uses it to make genre-defining titles. The studio’s first releases will target high-growth mobile and cross-platform markets, with gameplay experiences impossible to deliver without this AI-first foundation.

Piero captures it well in his GamesBeat interview:

“The magic happens when AI stops being a gimmick and becomes the canvas for the entire game.”

We’re hiring!

With this round, Studio Atelico will scale its engineering and design teams, ship the first public version of its AI engine, and launch its debut titles. We’re proud to back them from the start and can’t wait to play what comes next.

You can read more in the company’s funding announcement and GamesBeat coverage.